Norvin scores live calls for synthetic speech and hands your team an evidence-backed verdict in about six seconds — which models agreed, at which moments, and why. It recommends verification. It never blocks or terminates a call.
Commercial text-to-speech now passes casual human inspection. The channels where money and identity move — contact centers, third-party verification, telephone payments, help desks — have no instrument for the question "is this voice generated?" Every workflow that treats a convincing voice as evidence is quietly exposed.
Account-ownership changes, identity-recovery resets, payment authorization, SIM-swap and activation flows, TPV consent capture. The calls where a convincing voice is treated as proof.
Thresholds tuned on clean studio audio are measurably wrong on 8 kHz telephony. Norvin is calibrated per condition — clean and telco paths each get their own measured operating point.
Advisory verdicts, human override, per-tenant kill switch, full audit trail. A detection system that can prove what it does — and cannot act alone.
call level, 500 human-to-AI calls within a 1,000-call production evaluation, 95% CI [98.0, 99.7]; replicated at 98.2% [96.6, 99.1] on 500 fresh human-to-AI calls
listening-adjudicated at the ≥2-window rule on 500 production human-labeled calls, 95% CI [0.7, 3.7]
† Production and controlled evaluations with frozen operating points; full populations, sample sizes, and confidence intervals in the methodology. Field false-positive rates are listening-adjudicated; raw flags containing real machine speech are reported separately.
Listening adjudication uncovered screeners, voicemail, and recorded prompts inside the negative class. Why field false-positive claims fail without ears on the audio.
Read →RESEARCHA fresh, zero-overlap production cohort reproduced the fielded result while the raw human-label flag rate moved. The case for frozen operating points and adjudicated metrics.
Read →TECHNICALRemoving the third family destabilized the fused score geometry. Its value is distributional anchoring, not standalone discrimination.
Read →